AI as news interface

Conversation with a knowledgeable friend

A research tool from Google points to ways news organisations could be using AI as an interface rather than for content creation. Main points in the newsletter today:

• NotebookLM provides a prototype AI news interface

• A test-run of loading daily news into NotebookLM creates a competent news bot

• This “knowledgeable friend” is a hybrid news presenter/newspaper

• The model may be commercially defensible

Midjourney/Crawford Media

Google’s NotebookLM allows any user to upload multiple source documents to a private work zone and then query those documents using the Gemini Large Language Model (LLM).

This seems like a simple extension of using OpenAI’s ChatGPT to query user documents, but the details of how NotebookLM is implemented push it into new territory.

Up to 50 documents of up to half a million words each sit in the central data reservoir, waiting to be queried (25m words). This is an unprecedented tool for any researcher who needs both source depth and specificity. I have used it to load multiple historical documents from my town into one place. It immediately became a local history expert with a superhuman ability to retrieve information and provide references.

The system provides sample questions and document prompts, offering to make timelines, FAQs and study guides. Being an LLM, you can ask it anything at all. It occurred to me that this active dialogue with a defined cache of information suits daily news.

To test the idea out, I converted all the Guardian’s news articles above the fold* on Tuesday morning into PDFs and then loaded them into a Notebook. 

That was 16 stories. I then queried the system in a variety of ways.

I asked it to summarise the headlines.

That was boring, so I asked it what it thought the most interesting stories were. It produced a better result.

The most interesting news, in my opinion, is the convergence of several stories highlighting the complex and often fraught intersection of politics, individual beliefs, and social issues.

Melania Trump's revelation about her pro-choice stance, as reported in the sources, stands in stark contrast to her husband's policies and the general stance of the Republican party. This difference of opinion within a family so deeply embedded in American politics underscores how personal convictions can diverge from political affiliations.

Similarly, the sources' coverage of the backlash against the German language council's decision to relax rules around apostrophe use reveals how seemingly minor linguistic changes can evoke strong reactions, particularly when perceived as succumbing to the influence of a dominant culture.

In Australia, the ongoing turmoil within the Victorian Liberal party following the Moira Deeming defamation trial exemplifies the challenges of maintaining party unity amidst ideological differences and internal conflicts.

Then I tried:

• Asking it about the political orientation of the publisher. It refused to opine.

• To invent some stories in a similar vein. It refused that too.

• To write a TV bulletin script based on the news of the day.

Finally, I had a casual conversation with the AI about the day’s news.

Is this going to be a thing?

Any news organisation with any sense is currently blocking access to LLMs. The AIs can summarise, condense, and rewrite news effortlessly, in the process circumventing the current paltry protections of copyright law. NotebookLM demonstrates the power of the daily general news offering when offered up in the guise of a knowledgeable friend. The benefits:

• The news one organisation has sourced** is kept together as a proprietary whole.

• The limited knowledge base keeps the news conversation timely.

• The combination of proactive presentation and reactive query response is an interesting hybrid of print and broadcast news.

This agent-based interface, with the appropriate legal framework, could provide news organisations with a defensible product.

These are the current challenges:

• As with all current and “safe” LLMs, there is a tendency for NotebookLM to be boring.

• NotebookLM occasionally omits details or gets the wrong angle on a story.***

• To really work, this should be a spoken conversation. The current version of NotebookLM does not have this ability.

The killer combination would be to merge OpenAI’s incredible conversational ability (more on Advanced Voice Mode in coming weeks) with NotebookLMs propriety knowledge upload.

NotebookLM does have a nascent audio feature, but it’s a bit of a party trick at present. It will auto-generate a podcast featuring two hosts based on your material. This is mind boggling the first time you hear it, and almost mundane subsequently. I tried it with the daily news, and it produced a podcast where the hosts were shaking their non-existent heads over the folly of war and human suffering. It was pretty bad, and a very slow way to get news.

The model of a conversation with the user is more promising. The dynamic of asking questions suits news, echoing the centuries-old technique of headline and body copy. This is what radio and TV news has always lacked: the ability to cut to the chase by saying “tell me more about …” 

Implications

The above experiment reinforced the conclusion that news organisations must not allow external AIs to aggregate their content. They must use all means at their disposal - technical, legal, political - to prevent others using the information they unearth. The nexus between those who discover facts on the ground and those who bring this information to public notice must be maintained. If it is severed, then the system breaks.

This is not an easy problem to solve, but a conversational interface where the content is owned and controlled by news publishers could help.

Have a great weekend (nearly there),

Hal

* “Above the fold”: the portion of a website that is visible without scrolling on a desktop computer. A subjective and mostly outdated term. Still used by people like me.

** The need to define the news a publisher has actually created itself, as opposed to rewritten or republished, may by confronting. Traditional news uses other peoples’ stories a lot.

*** Example: Classifying the story of a veterinarian who refused to put down a cat as “a heartwarming story about a vet who followed her conscience”. That is not how I interpreted it. The vet was severely reprimanded after taking the sick cat home without the owner’s knowledge. The animal died a few months later.